The return to Australian fine wine

1. Introduction

The export success of the Australian wine industry has been well documented: in 2004 Australia exported wine worth almost US$ 2 billion, and is now the world's fourth largest wine exporter ( ABS, 2004 ). In fact, the only Australian agricultural commodities worth more in terms of export earnings to the Australian economy are wheat, beef and wool. Although Australian wine has increasingly displaced European wine as the table wine of choice in Anglophone countries, it remains to be seen whether Australian wine will also become the investment wine of choice in these countries.

Throughout the world, any discussion on wine investment inevitably turns to French wine, and in particular Bordeaux wine. The ability of fine Bordeaux wines to age majestically for decades, combined with the volume of production in Bordeaux, has ensured that the wines of Bordeaux dominate the trade in investment quality wines the world over. Yet Australia also produces many wines that improve in quality over time, and Australian Shiraz is particularly noted for its ability to improve with age. For example, certain vintages of Australia's premier wine, Penfolds Grange, not only improve in quality for decades after release, but also attract prices not dissimilar to those of many leading French wines. By providing estimates of the rate of return to premium Australian wine, this article provides insights into whether premium Australian wine represents a better investment option than Bordeaux wine.

The remainder of the article proceeds as follows. In Section 2, the previous literature concerning the rate of return to wine is reviewed, thereby establishing the general characteristics of the fine wine market, which is the context for this study. The hedonic price index methodology, which is used to estimate the return to premium Australian wine, is introduced and explained in Section 3. Section 4 outlines the dataset used and discusses the empirical results. Concluding comments are presented in Section 5.

2. Literature review

Table  1 presents a summary of the literature investigating the return to wine. Entries are listed in chronological order to show the evolution of knowledge in this area. The second column of the table, reporting estimation methods, indicates that historically the most popular method for estimating the return to wine has been the repeat sales regression approach due to Bailey et al . (1963) , although the hedonic approach, brought sharply back into focus by Rosen (1974) , has also been used. The third column of Table  1 shows that returns have been estimated from as early as 1969 to as late as 2002. The fourth column reports the type of wine included in each study and shows that the wines of interest have predominantly been the red wines of Bordeaux. Given the legendary ageing properties of wines from the leading chateaux in Bordeaux, and the historically dominant position of France in the wine trade, this focus on Bordeaux wine is understandable.

The final column of Table  1 summarises the key findings of each study, and from the details provided in this column it is possible to gain some insights into the behaviour of the wine market. The key features of the wine market appear to be that (i) returns to wine are both volatile and cyclical; (ii) external shocks, such as the introduction of a buyer's premium, may have substantial price effects; (iii) the return to wine over extended periods is likely to be higher than the return to risk‐free assets, but is probably not as high as for an equity portfolio; (iv) sub-market portfolios of particular vintages will outperform the return to a portfolio of wines from all vintages; and (v) returns may vary with price. Although the current study is concerned with premium Australian wine, it is reasonable to think the return to storing Australian wine will exhibit characteristics similar to those reported for wine from other regions of the world. At this point, it is worth noting that when calculating the return to wine, storage and transaction costs are important. For example, it is not uncommon for the buyer's premium at a wine auction to be 15 per cent. Unless otherwise stated, the return information presented in Table  1 and throughout the remainder of the article does not consider storage or transaction costs.

3. The hedonic approach

The hedonic approach to consumer demand analysis assumes that there exists a (hedonic) function relating the price of a good to the underlying attributes of the good. The literature using the hedonic approach to investigate the relationship between wine prices, objective wine characteristics, sensory characteristics and reputation attributes is both vast and growing, and some of the more notable contributions on the topic would include Combris et al . (1997) , Landon and Smith (1997) , Oczkowski (1994) and Schamel and Anderson (2003) . The hedonic approach has also been used successfully to look at the relationship between wine prices, quality and weather, and notable contributions on this topic include Ashenfelter et al . (1995) and Byron and Ashenfelter (1995) . Although various approaches could be used to estimate the return to wine, it seems natural to use the hedonic approach given that hedonic approaches have been widely used to study topics in wine economics. The approach, and some of the limitations of the approach, are explained below.

The separability assumption is strong, but is often made in consumer demand analysis. Furthermore, it may not be as troubling as once thought. For example, Clements et al . (1997) , using data for seven countries and employing a range of different tests, reported that preference independence, or strong separability, where sub-utility functions are explicitly assumed to be additive (a concept usually only employed when dealing with broad aggregates) cannot be rejected for goods as narrowly defined as beer, wine and spirits. The separability assumption is a limitation of the approach, but far from a terminal limitation.

Theory provides no guidance regarding the specific functional form of the hedonic price equation. Triplett (2004) argued strongly that functional form is a question to be answered by the data and the data alone. The implication of Triplett's argument is that the criteria used for selecting between different functional forms of the hedonic price equation (e.g. linear, semi-log or double-log) should be a series of Box–Cox tests. However, Diewert (2003) made an even stronger case for excluding the linear specification a priori . For functional forms using untransformed price as the dependent variable, it is not possible to satisfy either Fisher's time reversal test or the implied homogeneity condition, two desirable properties. When the logarithm of price is the dependent variable, these two properties are satisfied. Thus, it seems reasonable to rule out linear specifications of the hedonic price function. The semi-log specification is the one chosen in Section 4 to estimate the return to wine, and the Appendix contains a detailed discussion, based directly on Diewert (2003 : 21–25), showing the time reversal test and the implied homogeneity condition are satisfied for the semi-log specification of the adjacent period hedonic price model.

In recent times, several new hedonic approaches to estimating price change, including the continuously changing coefficient method ( Auer, 2004 ), and the generalised dummy variable approach ( Diewert, 2001 ) have been proposed. Yet despite this, the standard model for estimating hedonic indices remains the adjacent period approach, now generally attributed to Court (1939) . The adjacent period hedonic approach to estimating price change is a simple approach, but it remains an approach with much merit. It is explained as follows.

Let p_it denote the natural logarithm of the price of wine i ( i = 1,…,  n ) at time t ( t = 0,…,  T ) and let p_is denote the natural logarithm of the price of wine i at time s ( s = t − 1). We assume that each bottle of wine can be completely described by the underlying characteristic set { z_k }, ( k = 1,… ,  K ). Now, consider the two regression equations: p_is = β ₀ + ∑ _{k = 1} ^K β _kz_kis + u_is , and p_it = α + β ₀ + ∑ _{k = 1} ^K β _kz_kit + u_it , where the u_is and u_it are zero mean disturbance terms. The β ₀ and ∑ _{k = 1} ^K β _k coefficients, which are constrained to be the same for adjacent periods, control for the heterogeneity of the wines. Therefore, the α coefficient represents the average log price change between periods s and t , in a bottle of wine that is of constant quality. Specifically, and as outlined in Triplett (2004) , with log price as the dependent variable, the OLS estimate of α represents an index defined as the ratio of the unweighted geometric mean of quality adjusted wine prices in periods s and t . The α values can then be interpreted directly, or used to construct a price index series.

A major criticism of the adjacent period approach is that, for each pair of adjacent periods, the coefficients that control for quality variation are constrained to be the same. Certainly, if the gap between adjacent periods is large, say years, or the coefficients are constrained for multiple periods because of data limitations, then the restriction may not be appropriate. However, as the coefficients are only constrained in adjacent periods, when the adjacent periods are relatively short, say months or quarters, the restriction is not troubling. Furthermore, at least for cases where the adjacent periods are relatively close, there is empirical evidence suggesting the constraint of equality on the coefficients that control for quality in adjacent periods does not materially effect price index estimates.

The coefficients that control for quality variation are not constrained in adjacent periods when the characteristic hedonic price index approach is used. So, given the same dataset, if the constraint on the coefficients that control for quality variation in adjacent periods matters, an index based on the characteristics approach should vary significantly from an index based on the adjacent period approach. As shown in Triplett (2004 : 93), where summary information on a range of hedonic price studies, augmented in some cases by further details provided by the authors, is presented, the difference in price indexes based on the characteristics approach and the adjacent period approach, when the same dataset is used, is negligible.

4. Data and results

The data for the study were provided by Langton's auction house, and have been summarised at the quarterly frequency, covering the period 1989Q4 to 2000Q4. Although, there are other auctioneers of fine wine in Australia, the majority of sales take place through Langton's. Today Langton's use an electronic auction, however, for the sample period a sealed-bid auction method was used. Langton's recorded only when the auction took place, plus the highest and lowest sale price for each item. No data on the number of bottles in each lot was recorded. When compiling the data, if several lots of an identical wine were presented for sale at a particular auction, we took the unweighted arithmetic mean of the highest and lowest hammer price as the price. If more than one auction took place in any given quarter, the unweighted arithmetic mean of the prices recorded for each auction was taken as the sale price for that quarter.

The selection criterion for wines included in the study was straightforward. Regardless of variety, if the wine was listed in the Caillard and Langton (2001) classification of Australian investment quality wines, the wine was considered suitable for inclusion in the study. It is worth noting that the wines listed as of investment grade are not exclusively Cabernet based. As explained in Clarke and Rand (2001) , many grape varieties, when treated appropriately, age gracefully for decades. In Australia, there are no formal restrictions with respect to which grape varieties can be planted in which region, so restricting the study to just Cabernet-based wines is not appropriate in the Australian context. In total, the sample comprised 14,102 observations relating to 84 different wine brands. The varieties represented in the sample and the number of observations on each variety are Shiraz (5,356), Cabernet (6,086), Merlot (30), Pinot Noir (820), Botrytis (110), Chardonnay (1,192), Semillon (183) and Riesling (325). Vintages prior to 1965 were excluded from the sample as they are likely to be traded as antiques.

Figure  1 summarises some characteristics of the price dataset. Panel A of Figure  1 shows the number of price observations in each quarter, and illustrates the increase in turnover during the sample period. Panel B of Figure  1 shows the number of observations per vintage and illustrates an intriguing aspect of the wine auction market. Two of the most celebrated recent Australian vintages have been 1986 and 1990. By contrast, vintage 1989 was a difficult vintage, and the wines from this vintage are generally considered to be of below average quality. Panel B of Figure  1 suggests that wines from poor vintages disappear from the market faster than wines of average or high quality. Such an interpretation is consistent with the findings previously noted for Bordeaux wine in Ashenfelter et al . (1995 : 9) and Burton and Jacobsen (2001 : 342).

The hedonic price approach assumes it is possible to completely describe the underlying attributes of the product in question. If the hedonic regression suffers from an omitted variables problem, then in general the estimates of price change will be both biased and inconsistent. Therefore, a comprehensive specification for the hedonic price equation has been chosen. The study assumes the price of a bottle of wine sold at auction is completely described by (i) time of sale (1989Q4–2000Q4); (ii) vintage (1965–2000); and (iii) a brand variable. There are 84 different wine brands in the study, so alongside the 35 vintage dummy variables, there are 83 different brand dummy variables. The base vintage is always the earliest vintage (not always vintage 1965), and the base brand is always Penfolds Grange.

As reported in Table  1 , there is some evidence that the risk and return to wine vary with price. Jaeger (1981) , investigating red Bordeaux and California Cabernet wines, found the least expensive wines had both higher returns and higher risk. This result has an intuitively appealing explanation. Less expensive wines are traded less frequently than more expensive wines, and so exhibit greater risk. To compensate for this extra risk, less expensive wines have higher returns than the more frequently traded, less risky, expensive wines. Burton and Jacobsen (2001) also found that the return to wine varied with price. Specifically, they found the most expensive wines had lower returns and higher risk than the aggregate market portfolio. However, it should be noted that as long as individual wine returns are not perfectly correlated, it is reasonable to expect that any given sub-market wine portfolio will exhibit more risk than the market portfolio. We should therefore not read too much into the finding that a portfolio containing only expensive wines had higher risk than the market portfolio. Although the risk-return profile of a portfolio of less expensive wine is not reported, the information that is reported in Burton and Jacobsen (2001) is not inconsistent with the idea that more expensive wine should have both lower risk, and lower returns, compared to less expensive wine.

Given the possibility that returns vary with price, the return to wine was also estimated for two sub-samples, expensive wine and less expensive wine. The Caillard and Langton (2001) classification of investment quality Australian wine allocates wines into four categories. The categories are exceptional , outstanding , excellent and distinguished . The most expensive wines are generally those with the rating exceptional, the next most expensive generally have the rating outstanding, and so on. In this study, the more expensive wine sample includes all wines with the rating exceptional or outstanding, and the less expensive wine sample includes all wines with the rating excellent or distinguished. For both sub-samples, the base vintage is always the earliest vintage. For the more expensive wine sub-sample, the base brand is Penfolds Grange, and for the less expensive wine sub-sample, the base brand is Wynns Coonawarra Estate Cabernet Sauvignon. ¹

As there are three specifications, and the sample period runs from 1989Q4 to 2000Q4, there are (44×3 = 132) separate adjacent period hedonic price regressions. Space constraints do not allow us to show the complete estimation results. Table  2 presents details on the estimated quarterly price change for the three model specifications, along with summary goodness-of-fit information. The parameter ᾱ estimates the log price change for the quarter. By considering the first row of the table, it can be seen that the estimated log price change for all wine in the first quarter of 1990 was −0.009. Table  2 also shows the associated percentage return to wine, which for 1990Q1 is (exp (−0.009) − 1)×100 = −0.88 percent. Considering the more expensive and the less expensive wines separately, it can be seen that for 1990Q1 the estimated log price change was −0.025 for the more expensive wine, and 0.009 for the less expensive wine. The return to more expensive wine for this quarter was therefore −2.51 percent, whereas for less expensive wine it was 0.93 percent. The R² values for each regression are also shown in Table  2 . If appropriate, the standard errors reported in Table  2 were corrected for heteroscedasticity.

From the information provided in Table  2 , the average risk and return information for the three specifications were calculated. ² The estimated mean quarterly return to all wine over the sample period was 2.35 per cent, with standard deviation 4.42 per cent. The results for the sub-samples are interesting. The average of the estimated quarterly return to the most expensive Australian wines was 3.17 per cent, with standard deviation 4.74 per cent, and the mean R² across all 44 expensive wine adjacent period regressions was 0.92. For the less expensive, although still premium, Australian wine, the average mean quarterly return was 1.92 per cent, with standard deviation 5.35 per cent, and mean R² of 0.76. As can be seen by comparing the risk associated with the three portfolios, the risk associated with the market portfolio is less than the risk associated with either wine sub-market portfolio. This result simply indicates that the returns to the most and the less expensive wines are not perfectly correlated.

That more expensive wine provides a higher return with lower risk than less expensive wine is an unexpected finding. In Australia, more expensive wine trades at auction with greater frequency than less expensive wine. For example, the seven wine brands in the study with the auction rating exceptional account for 22 per cent of all sale observations. As less expensive wine is traded less frequently, the standard deviation of returns to less expensive wine should be, and was in fact found to be, greater than for more expensive wine. Ex ante we would therefore expect less expensive wine to provide a higher return to compensate for the higher risk. However, the fact that this expectation is not realised ex post here, simply means that the expected volume of buyers of less expensive wines did not materialise. Furthermore, it is worth remembering that the secondary market for wine in Australia is young. In such an undeveloped market, deviations between ex ante expectations and ex post results are always a very real possibility.

Although the estimated coefficients attached to the brand dummy variables are not reported, these dummy variables play an important role in the hedonic regression specification. The brand coefficients show differences relative to the base brand, which for the ‘all wine’ regressions is Penfolds Grange. As Penfolds Grange is noticeably more expensive than almost all other wine in the sample, the coefficients of almost all the brand dummy variables are statistically significant at the 5 per cent level in each adjacent period all wine regression. A more appropriate indicator of the important role played by the brand dummy variables can be obtained if the approach for interpreting dummy variables due to Suits (1984) is used. Using the Suits (1984) approach, standard t -tests on the brand coefficient estimates test the hypothesis that a particular brand effect is statistically different from the mean brand effect. Taking the last adjacent period regression for the all wine sample as a representative case, and using the Suits (1984) approach, t -tests indicate that, at the 5 per cent level, 76 out of the 84 brand effects are individually statistically different from the mean brand effect. The extent of the role played by the brand dummy variables can also be seen by considering the range of dollar values implied by the estimated coefficients. Again, taking the last adjacent period regression for the all-wine sample as a representative case, the wine with the largest brand effect had an implied price more than four times the average price, whereas the wine with the smallest brand effect had an implied price of less than half the average price.

Placing the estimated wine return information in context is also a worthwhile exercise. Table  3 compares the return to wine with the return to other financial assets for the same period. The share price returns have been calculated from an Australian All Ordinaries total return index, and so capture both the change in the capital value of Australian equities and dividend payments. The coefficients of variation in the final column of the table are a measure of risk per unit of return, and are calculated as the standard deviation of returns divided by the mean return. Hence, in the Australian context, in terms of returns, equities (quarterly return 2.67 per cent) appear to dominate wine (quarterly return 2.35 per cent). However, in terms of risk per unit of return, wine, with a coefficient of variation of 1.88, has a slight edge over equities, whose coefficient of variation is 2.18.

As noted in the literature review, the return to wine has historically been cyclical. It is therefore not wise to make return to wine comparisons across markets at different points in time. There is, however, a seven-year overlapping period between the current study and the Burton and Jacobsen (2001) study of red Bordeaux wines. Using the data reported in Table  1 of Burton and Jacobsen (2001 : 343), we calculated that the average semi-annual rate of return to red Bordeaux wine for the period 1990–1996 was 4.06 percent (standard deviation 12.39 per cent, coefficient of variation 3.05). The mean quarterly rate of return to premium Australian wine for the period 1990–1996 was 2.50 per cent (standard deviation 5.09 percent, coefficient of variation 2.04). In general, Australian wine sold at auction is both less frequently traded than Bordeaux wine, and less expensive. It is therefore not altogether surprising that Australian wine should provide higher returns than Bordeaux wine. However, that these higher returns are not associated with higher risk is something of a surprise, and perhaps reflects a lack of published information concerning the rate of return to Australian wine.

5. Conclusion

The Australian wine industry has grown substantially in recent decades, and the export success of the Australian wine industry is widely acknowledged. At the same time as the retail wine industry has been growing strongly, so too has the Australian wine investment market. To date, the economics of wine investment in Australia have not been well understood. This article, by using hedonic price techniques to estimate the return to wine, has revealed much about the investment properties of Australian wine. The results presented suggest that, among premium Australian wines, more expensive wine is less risky than less expensive wine, the return to more expensive wine is higher than the return to less expensive wine, and the risk-return profile of wine is broadly comparable to the risk-return profile of Australian equities. The return to Australian wine was also compared to that of Bordeaux wine for a seven-year overlapping period. During that period, Australian wine offered higher returns for less risk than Bordeaux wine, although such a short overlapping period cannot provide conclusive evidence.

Appendix (based directly on Diewert, 2003 : 21–25)

It is desirable that the hedonic price equation used to estimate the return to wine be linearly homogeneous. Using OLS estimation and the adjacent period hedonic price regression approach, homogeneity may be thought of as implying the following condition: if all n wines are sold in period s and period t , and all n wines increase in value between the two periods by λ, where λ is some positive constant, then the OLS predicted values in period t should be exactly λ times the period s predicted values.

Taking the antilog of both sides of the above equations gives P_is = exp [β ₀ + ∑ _{k = 1} ^K β _kz_kis ]exp[ u_is ] and P_it = exp[α _st ]exp[β ₀ + ∑ _{k = 1} ^K β _kz_kit ]exp[ u_it ]. When estimated using OLS, the implied homogeneity condition that the predicted values in period t be exactly λ times the period s predicted values requires exp[ᾱ _st ]exp[β̂ ₀ + ∑ _{k = 1} ^K β̂ _kz_kit ] = λ exp[β̂ ₀ + ∑ _{k = 1} ^K β̂ _kz_kis ], a condition met when λ = exp[ᾱ _st ]. Hence, in linear adjacent period hedonic price models, when the natural logarithm of price is used as the dependent variable, the implied homogeneity condition is satisfied. Unless equation (A2) is written in nonlinear form, this property does not hold when price, untransformed, is used as the dependent variable.

Now, let the data be ordered so that all observations from period s appear before the observations from period t . Let denote a column vector of length equal to the number of observations from period s , all the elements of which are zeros, and let denote a column vector of length equal to the number of observations in period t , all the elements of which are ones. Given this notation and ordering, the relationship between the OLS parameter estimates of equation sets (A3) and (A4) is: , which implies the following two equations: and . Equations which in turn simplify to: and , which imply ᾱ _ts = −ᾱ _st . Taking the antilog of both sides gives exp[ᾱ _ts ] = exp[−ᾱ _st ] and so I_ts = 1/ I_st . With log price as the dependent variable the time reversal property is satisfied. This is not true when price untransformed is used as the dependant variable as the equations imply I_ts = − I_st .

Acknowledgements

The author would like to acknowledge the many helpful comments of Ken Clements, and three anonymous referees. The comments provided by one referee with respect to the hedonic model specification were especially helpful.

References